Image segmentation is a commonly performed clinical practice for extracting geometrical descriptions of internal anatomical objects from volume images such as computed tomography and magnetic resonance images. Shortcomings of current practice have motivated an important area of research involving statistically trainable deformable-shape models (DSMs) for automatic segmentation. The general approach is to apply a DSM to an image and cause the model to undergo a series of deformations converging to a close match between the model and the target anatomical object(s). The deformations are driven, in an appropriate statistical framework, by mathematical optimization. The overall aim is to develop and clinically evaluate a workstation for automatic segmentation of anatomical structures in the male pelvis from CT images for image- guided applications in radiation therapy using technology based on a particularly powerful class of DSMs called m-reps. Segmented CT images provide guidance for critical treatment planning and radiation delivery decisions in radiation therapy. It is likely that segmentation is performed more often for these purposes than for all the other medical applications combined. Current interactive contouring methods in clinical practice are extremely time consuming and expensive, and the contours demonstrate significant inter- and intra-user variabilities that adversely affect the clinical decisions that rely on them. The proposed methodology will overcome these shortcomings and improve the effectiveness of radiation therapy for treating cancer.